An image from Can overfitted deep neural networks in adversarial training generalize? – An approximation viewpoint, one of the productions that also features Fanghui Liu.